Uncertainty Reduction in Diffusion Magnetic Resonance Imaging Tractography

Abstract

Diffusion Magnetic Resonance Imaging (dMRI) is currently the only non-invasive method capable of mapping the geometry and microstructure of major white matter tracts in vivo. This technique measures the movement of water molecules along magnetic field gradients. Since fiber tracts impede water movement perpendicular while facilitating it along their length, dMRI can measure intricate microstructural information. <br /> Tractography, the process of reconstructing streamlines representing fiber path bundles, has become indispensable in brain studies and surgical planning due to its low risk and high image quality. However, the entire tractography pipeline, from dMRI measurement through pre-processing to final tractography, is highly susceptible to uncertainties, stemming from complex measurement schemes and model imperfections. <br /> In this work, we study multiple sources of uncertainty in tractography and propose models to mitigate these challenges. Within our first contribution, model uncertainties, stemming from the necessary choice of an appropriate model a-priori, and measurement uncertainties are examined. Further, the interaction between both types of uncertainties is investigated and novel methods to reduce them are introduced. It is shown that reducing model uncertainty also reduces the susceptibility to measurement noise. Finally, the impact of both methods to tractography is discussed and it is visually demonstrated that both methods increase the completeness compared to a more restrictive model selection approach, while at the same time reducing false positives, compared to the low-rank approach. Quantitative analysis over several subjects from the Human Connectome Project (HCP) and a variety of tracts supports the visual impression. <br /> Within our second contribution, we introduce two novel regularized tractography methods to stabilize the tractography against errors in the local direction field, which are amplified if not taken into account. The first method computes a joint low-rank approximation by considering the local neighborhood, while the second employs an unscented Kalman Filter (UKF) to update the local configuration in each step, based on previous estimations. The approaches are evaluated on several datasets, such as a high quality dataset from the HCP, a challenge dataset for tractography and a clinical dataset from a tumor patient. Results show that both regularization techniques lead to a significant improvement of reconstruction quality, compared to their unregularized counterparts, Pareto optimal results on the challenge dataset, and also high quality reconstructions within measurements of a tumor patient. <br /> However, the low-rank UKF model, while effectively regularizing, occasionally misses some parts of fanning tracts. This originates in the low-rank model, which is able to recover fiber crossings but lags in accounting for fanning. Within our third contribution, we have combined the Bingham distribution, which is able to capture fanning on the sphere, with the low-rank model. Implementing this model involved to overcome several technical challenges. The approach is evaluated on the HCP dataset for a variety of tracts. Results indicate its clear superiority over the previously presented low-rank UKF model in terms of capturing fanning, while maintaining specificity.